Why Government IT Projects Fail: HealthCare.gov, the FBI's Virtual Case File and the Lessons From Phoenix, Horizon and Robodebt

On 1 October 2013 HealthCare.gov opened and buckled on day one. The HHS inspector general later counted nearly $800 million obligated across 60 contracts by February 2014. A decade earlier the FBI had scrapped its Virtual Case File after three years and $170 million. And GAO says the federal government still spends about 80% of its more than $100 billion a year in IT on running systems it already has.

The US isn’t unusual in this. Canada’s Phoenix pay system shows the same failure at its most extreme. Canada approved a new federal pay system in July 2009. The budget was about $310 million, and the promise was savings of $70 million or more a year. Phoenix went live in 2016. By June 2025 a Public Services and Procurement Canada official put the cost of dealing with its errors at more than $5.1 billion, and it still isn’t fixed.

The Auditor General’s report tabled in March 2026 counted a backlog of 233,653 pay transactions at the end of September 2025, affecting 133,619 public servants. That’s half the peak of 447,080. Ottawa is now building a replacement on Dayforce, estimated at more than $4.2 billion. So a project meant to save money has cost more than sixteen times its first budget just to clean up after, plus the price of a second system to replace it.

Phoenix is extreme. It isn’t unusual.

The same failure, in four countries

Britain’s National Programme for IT was supposed to give the NHS a single electronic patient record. The government announced in September 2011 that it was dismantling the programme. Two years later the Public Accounts Committee put the forecast cost at £9.8 billion and called it one of the worst contracting fiascos it had seen.

In the US, HealthCare.gov launched on 1 October 2013 and buckled on day one. The HHS inspector general later found nearly $800 million obligated across 60 contracts by February 2014. Its 2016 case study listed the causes plainly: no clear leader, policy work that crowded out the build, poor technical choices and a culture that didn’t want to hear bad news.

The FBI had already been there. It scrapped its Virtual Case File in 2005 after three years and $170 million, then started over with Sentinel at an estimated $425 million.

Australia’s Robodebt is a different kind of failure, and a darker one. The software worked as designed. The design was the problem: it averaged annual tax data to claim welfare overpayments that often didn’t exist. A Royal Commission reported in July 2023 with 57 recommendations. In March 2026 the National Anti-Corruption Commission found two senior officials had engaged in corrupt conduct. In June 2026 the Federal Court approved a further A$548.5 million settlement, on top of an earlier A$112 million. About 450,000 people were caught up in it.

Then there’s Horizon, the Post Office accounting system that wrongly showed shortfalls in branch accounts, which led to hundreds of sub-postmasters being prosecuted. The public inquiry published its first volume, on human impact and redress, in July 2025. Volumes two to six are still waiting.

What the pattern looks like

Read the post-mortems side by side and the technology is rarely the story. Five things keep coming back.

The scope is set by policy, then handed to IT as fixed. HealthCare.gov had a political launch date that didn’t move, whatever state the system was in. Phoenix was approved on a savings case, so the savings became the goal and the pay system became the means.

Nobody owns the whole thing. Big programmes split responsibility between a policy department, a delivery agency and a prime contractor. When it breaks, each can point at the other two.

Bad news travels slowly upward. Almost every inquiry finds staff who knew, early, and a hierarchy that didn’t want to know.

Contracts reward delivery of the spec, even when the spec is wrong. The contractor builds what was ordered. The government finds out what it actually needed at go-live.

And there’s no way back. Old systems get switched off too early, so a failing new one can’t be paused. Once the old payroll was gone, Canada had no choice but to keep paying people through Phoenix and fix it live.

The quieter problem is the old code

Failed launches get the headlines. The bigger cost sits in systems nobody is replacing at all.

GAO says the US federal government spends more than $100 billion a year on IT, and about 80% of that typically goes to running and maintaining what already exists. In 2019 GAO flagged ten critical legacy systems. By February 2025 only three had been modernised. The eleven systems on its current list are between 23 and 60 years old.

IT acquisition has been on GAO’s high-risk list since 2015. In the 2025 update it was one of three areas that went backwards. GAO has made 1,881 recommendations on it since 2010, and 463 were still open in January 2025.

That’s the trap. Legacy systems eat the budget, so modernisation happens in big, rare, high-stakes leaps, which are exactly the projects that fail. Countries that handle this better tend to replace systems in smaller pieces, on shared components.

None of the four countries above lacked money or talent. They lacked a way to stop. The one lesson every inquiry agrees on is the cheapest to apply: build in a point where someone can say the system isn’t ready, and have it count.